Sfx

• /sfx "<prompt>" — minimum form. Defaults: 2s, no loop, apps/web/public/sfx/misc/, mp3_44100_128.
• /sfx <class> "<prompt>" — class drives the output subdirectory. Valid: ui, foley, cinematic, mechanical, ambience, event, misc. Path becomes apps/web/public/sfx/<class>/.
• /sfx <class> <name> "<prompt>" — name becomes the filename stem (<name>.mp3). Without a name, the MCP picks one.
• /sfx <class> <name> <duration>s "<prompt>" — explicit duration in seconds (0.5–5).
• /sfx <class> <name> <duration>s loop "<prompt>" — loop ON.

If invoked with no prompt at all, ask the user for one or hand off to the sound-design skill to author one.

If the ElevenLabs MCP tools are not available in the current session, stop and tell the user this environment cannot generate the audio directly. Do not fake a generated file.

1. Validate

• Prompt length: warn if < 10 words (likely too generic) or > 60 words (likely contradictions). Don't block — warn and proceed.
• Duration: clamp to [0.5, 5]. The MCP rejects values outside this range. If the user requested > 5s, tell them and proceed at 5s, OR ask if they want to chain calls / switch to the public UI.
• Loop sanity check: if loop: true and the prompt looks event-shaped (contains "slam", "click", "hit", "impact", "shot", "drop", "snap"), warn the user — looping makes no sense for transient events. Ask before flipping it off.
• Multi-event sanity check: if the prompt contains commas-then-conjunctions like "then", "followed by", "and then", "after that" implying a sequence of distinct events, stop and route to the sound-design skill — the consolidated framework's load-bearing rule is "one event per generation".

2. Resolve the output path

• Class → subdirectory: apps/web/public/sfx/<class>/ (default misc).
• Create the directory if it doesn't exist (mkdir -p).
• Filename: if the user provided one, use <name>.mp3 (or matching extension for chosen output_format). If not, let the MCP name it.
• Never save outside apps/web/public/sfx/ without explicit user override — that directory is the project convention for static SFX served by Vite.
• Confirm the path back to the user before generating, so they can redirect if needed.

3. Call the MCP

mcp__elevenlabs__text_to_sound_effects(
  text:              "<the six-slot prompt>",
  duration_seconds:  <0.5–5, default 2>,
  loop:              <true|false, default false>,
  output_directory:  "<absolute path to apps/web/public/sfx/<class>/>",
  output_format:     "mp3_44100_128",  // override only on user request
)

This is a paid API call. The MCP tool's own description carries a cost warning. Each call generates ONE file (the MCP returns one variation, not the four the public UI returns). If the user might iterate, warn them up front: "I'll generate one variation. If it misses, we re-prompt — each call has a credit cost."

4. After the MCP returns

• Surface the saved file path to the user — exact, copyable.
• Show the prompt, duration, and loop flag that produced it (so the user has the recipe).
• DO NOT auto-play. If the user asks, call mcp__elevenlabs__play_audio with the returned path.
• DO NOT auto-git add. Tell the user the file is on disk; they decide whether to track it.

5. If iteration is needed

If the user is dissatisfied, do NOT immediately call /sfx again with the same prompt. Hand off to the sound-design skill which owns the iteration protocol (change one slot, simplify, add naturalism cues, etc. — see docs/sound-design/CONSOLIDATED.md). One regen with the same prompt is a coin flip; the framework gets you to "right" in 2–3 deliberate iterations.

SAVED:    <absolute path>
DURATION: <Ns>
LOOP:     <true|false>
PROMPT:   "<the prompt that produced it>"

Play it?  Say "play" — I'll call mcp__elevenlabs__play_audio.
Iterate?  Tell me what's off; I'll route to the sound-design skill for round 2.